I build software in Wollongong: fast local AI on Intel Arc GPUs, planning tools, and a train game.
- llama.cpp on Intel Arc (SYCL): XMX flash attention for quantized KV caches (up to 2.6x faster generation at long context on a B580), XMX prompt GEMMs for the default fp32 build (~3x faster prompt processing), and a faster GATED_DELTA_NET kernel for Qwen3.5-style models. Sent to Intel's SYCL maintainer for upstreaming. Branches: sycl-fa-q4-xmx, sycl-xmx-prompt-gemm, sycl-gdn-blocked
- Bonsai 2 27B on Intel Arc B580: SYCL kernels that run PrismML's ternary 27B model on a 12 GB Arc B580 with the full 128K context, ~85 t/s with speculative decoding.
- TernSYCL (contributor, 4 merged PRs): int4/int8 DPAS helpers and fixes for ternary GEMM kernels on Xe GPUs.
- wayfinder: lets an AI agent point at things on your Windows screen instead of clicking for you.
- Next Stop NSW: a browser train game built on the real NSW rail network.
- AusDatum: NSW site and planning analysis: zoning, hazards and development pathways for any lot.
- PyAtlasPro: a free pyRevit extension for faster Revit drafting.
Best way to reach me is LinkedIn.

